Keywords

1 Introduction

The human body contains different components, which function with several physiological processes. It has been seen that most of these physiological processes represent themselves in the form of signals and these signals are of different types, viz. biochemical, biomedical, biomechanical, etc. Generally, biomedical signals hold a lot of information about the changes in the physiological activities of our bodies [1,2,3]. Disease and ailments may create a disturbance in the biological system of humans which causes disturbance in the smooth functioning of our daily activities. It has been observed that these diseases and ailments in the human biological system are counter-productive leading to several pathological conditions which affect the health and well-being of the system [4,5,6]. Analyzing these biomedical signals was quite difficult a few years ago since the obtained signals were contaminated with different types of irrelevant noise and powerline interferences. Sometimes it is also required to obtain their spectral properties to understand the characteristics of the signals’ behavior in terms of the different frequency ranges, and modeling of the different feature representation and parameterization. The other major challenge concerning biomedical signal processing is the experience and expertise of the clinicians in analyzing and interpreting the signals in a proper way [7].

However, this entire process is subjective. In recent research developments, digital signal processing and pattern recognition have been the backbone of biomedical research applications [8, 9]. Another crucial aid to biomedical research is a computer-aided diagnostic system which assists the domain experts to make important decisions during the diagnostic process plus it also offers the possibility of online monitoring for critically ill patients. Thus, there is a huge demand for more reliable, effective, and affordable clinical and health services. Hence there are so many novel medical equipment, techniques being developed to support such diagnosis, monitoring, and treatment of irregularities, abnormalities, and ailments in the human body. In the present world, one of such major health issues has been sleeping deprivation and associated illnesses which adversely affect the quality of life across the global population irrespective of age group. Hence applications based on sleep studies and patterns become crucial to diagnosing and proposing solutions to such ailments. EEG signals are most appropriate to identify irregularities in the regular sleep pattern more effectively. Additionally, it also has some more advantages that are it is high speed, time-efficient, and also non-invasive. Therefore, the diagnosis with EEG signals is inexpensive and widely accepted in the diagnosis of different sleep-related diseases [10]. It has been seen from the recent research developments in the signal processing techniques, an improved approach for classification on the basis of PSG signals has been reported. These enhanced techniques can help to recognize the abnormal sleep patterns and irregularities in sleep leading to sleep-related diseases. An automatic sleep staging system is obtained in the diagnosis of several sleep disorders, which is considered one of the best tools apart from the visual inspection of the EEG signals by physicians [11].

1.1 Related Work

Some of the latest works on automated sleep stage classification making use of single-channel and multiple channels are presented in Table 1.

Table 1. A brief analysis of state-of-the-art works for the sleep staging
Full size table

From the state-of-the-art works, it can be seen that most of the studies have focused on single-channel and multiple-channel. Some of the studies were well performed with the input of single-channel, additionally, it gives well comfortable for the subject during sleep recordings. This advantage gives an improvement in sleep staging accuracy. But it has been also observed that major of the studies highly suffered from data imbalance problems, misclassification of the N1 sleep stage, and improper selection of features for the classifier.

Therefore our proposed work has different from other similar works in this context; we have obtained two different session recordings for the subject who was affected with a sleep disorder. This work aims to contribute to the comparative analysis of sleep staging methods by assessing the classification accuracy of statistical moments derived from single-channel EEG signals. With regards to classification techniques, we used SVM for testing them between participating subjects in the research work. We report all Cohen’s kappa values for various sleep stages for the comparison of the classification algorithm. The proposed algorithm can be observed to present high values of accuracy and Cohen’s kappa of the data of healthy subjects.

Further, our proposed research work is organized as follows: Sect. 2 presents the data used in the experimental work. In Sect. 3 we briefly present the proposed methodology. Section 4 briefly about the outcome of the experimental analysis. Section 5 remarks the concluding remarks.

2 Experimental Data

In this study, for experimental analysis, we have acquired two different categories of the patient’s records, one from the subjects who were already suffered from mild sleep-related problems and the other category of the subjects, who were completely healthy controlled. The whole dataset is an accumulation of three types of sleep recordings belonging to different patients. The entire recordings were collected by a group of expert clinicians [25]. The whole recordings were obtained from different patients, who were already faced various kinds of sleep-related problems, and the other part of the recordings were obtained from healthy subjects.

In this work, for our experimental study, we used the C3-A2 channel for computing sleep stage classification. Table 2 presents a detailed account of the percentage of sleep epoch distribution in different sleep stages.

Table 2. Description of the percentage of epoch per individual sleep stages
Full size table

This study includes three categories of subjects. One is affected by sleep problems with one session recording sleep data. The second category of data was extracted from subjects with sleep disorders in two sessions of recording. Finally, the third category includes healthy subjects. The authors have obtained the EEG channel of C3-A2 signals of 6 subjects. The stage annotations are also provided in the data repository according to AASM rules. The unscored epochs are not considered for further analysis in the experimental work. Each epoch is considered a 30 s time length in this present work. For analysis, of the sleep behavior, we have presented the samples of all sleep stages of EEG signals extracted from different categories of subjects. The sleep EEG signals behavior changes according to the sleep cycle covered by the subject. Every stage of sleep is characterized by different behavior from sleeping brains such as low amplitude, mixed frequency, sawtooth waveforms, low amplitude in muscle movements, and rapid eye movements. However, the EEG signal behaviors are complex since they are not periodic and also because of the continuous changes of amplitude, frequency, and phase range according to the sleep stage. In the proposed method, the authors have applied the Butterworth bandpass filter to reduce the undesired segments from the input signal. The filter is a second-order Butterworth bandpass filter, and the frequency ranges applied range from 0.5 Hz to 35 Hz.

3 Methodology

This study demonstrates a unique method of two-stage sleep scoring using a single EEG channel. Figure 1 represents the workflow of the proposed sleep stage classification system.

3.1 Proposed Automatic Sleep Stage Detection Method

The sleeping study can be divided into four layers. These are the subject layer, classification layer, section layer, and instance layer. The subject layer is divided into two sections. One section contains information about EEG, and another section represents sleep information. These two sections are the main domains of this study and compose the classification layer. The section layer represents the associated concepts related to the EEG signals and sleep sections. While the EEG section defines the used electrode name, obtained features, and participants enrolled for this experimental work, the sleep section addresses the obtained sleep stage rules for sleep quality analysis. Finally, the instance layer considered the specific settings associated used the experiments conducted in the research and is defined according to the other layers. In this proposed study, two different medical condition subjects’ details from the ISRUC-Sleep dataset from the University Hospital of Coimbra, Portugal are evaluated. We have made use of the OSFS systems for feature selection [26, 27]. The final feature selection output is presented in Table 3.

Fig. 1.
figure 1

Workflow of the proposed method

Full size image

Algorithm 1

Automated SleepEEG Classification Test.

Procedure: Sleepstage_Scoring = SleepStage_Classify (Feature List, Labels).

Input: Patient Records [Extracted from Sleep_Dataset]

$${P}_{V}=\left\{{P}_{1},{P}_{2},\cdots \, \cdots \, \cdots \, \cdots,{P}_{N}\right\}$$
$${P}_{SI}=\left\{{S}_{1},{S}_{2}, \cdots \, \cdots \, \cdots \, \cdots \, \cdots,{S}_{M}\right\}$$
$${P}_{SE}=\left\{{S}_{I1},{S}_{I2}, \cdots \, \cdots \, \cdots \, \cdots \, \cdots, {S}_{IZ}\right\}$$

%N = Number of Patients enrolled for SleepEEG test.

%M = Number of Session recordings considered per individual enrolled subjects.

%Z = Number of Sleep stages epochs.

Pre_Predict_Label = pre-predicted annotations of sleep stages.

Output: Sleepstage_Scoring = Correct predicted sleep stage sequences.

Step-1: Let P be the enrolled patient’s records from different health conditions.

Step-2: Let K be the total sleep recordings of enrolled subjects from a single channel of EEG signal.

Step-3: For each EEG recording from individual enrolled patients in the SleepEEG study do.

Step-4: P = P + 1.

Step-5: Divide the whole K segments of sleep recordings into 30s epochs-wise.

Step-6: Extracting both time and frequency domain features from each epoch and stored into specific feature vectors per individual enrolled patients.

figure a

Step-7: Forwarded the extracted features to the proposed feature selection techniques and select the suitable features and store them into the selected feature vector patients-wise.

figure b

Step-8: Forwarded selected feature vector list and predicted sleep stage segments into the proposed classification model.

Sleepstage_Scoring = SleepStage_Classify (Feature_Selection_PN, Pre_Predict_Label).

Step-9: if the Sleep stage classification accuracy is suitable then recommend the sleep stage scoring approach in the diagnosis of sleep-related problems, go to step 11.

Step-10: else go to step 6 [For new feature extraction from sleep recordings of the enrolled subjects].

Step-11: end if.

Output: Automated Corrected Sleep Stage Sequences of the Subjects.

Table 3. Final feature selection list
Full size table

4 Experimental Results and Discussion

This complete methodology procedure of this proposed work is explained in pseudo-code format, which is described in Algorithm 1. The complete illustrations of the experimental steps of this proposed work are shown in Fig. 1. For each sampled patient, the experimental results were concluded separately. Multiple evaluation metrics such as accuracy (Acc) [28], recall (Rec) [29], specificity (Spe) [30], precision (Pre) [31], F1score (F1sc) [32], and Cohen’s Kappa Score [33, 34] to analyze the performance of the proposed sleep studies.

4.1 Classification Accuracy of Category-I Subjects ISRUC-Sleep Database

Table 4. Performance evaluation results for ISRUC-Sleep subgroup-I
Full size table

The results for subject-1, subject-2, subject-3, subject-4, subject-5, subject-6, and subject-16 show an overall classification accuracy of 91.2%, 94%, 92.4%, 97.73%, 93.86%, 93.46%, 95.60% and 91.46% using SVM classifier. The higher precision value is presented for subject-4 using the SVM method of 98.63%. The highest specificity and F1-score performance results are 92.24% (Subject-5) and 98.83% (Subject-4) using the SVM method. The Kappa scores for subjects-2, 5, and 16 are found in excellent agreement with the subject to the best accuracy for investigation of sleep irregularities. The reported evaluation metrics performances for all subjects for both the session recordings are described in Table 4.

4.2 Classification Accuracy of Category-III Subjects ISRUC-Sleep Database

Finally, the final experimental results for the proposed SleepEEG method are done through the analysis of signals taken from healthy subjects, who haven’t been on medication of any sort prior to the experiment. The one-session recording was taken into consideration for monitoring the sleep irregularities and the performances achieved for all the healthy controlled subjects are presented in Table 5.

Table 5. Performance evaluation with ISRUC-Sleep subgroup-III data
Full size table

It can thus be observed that the model achieves an overall accuracy of 92.5%, 92.26%, 91.73%, 91.6%, 96.53%, 66.13%, 95.06% and 93.06% through SVM classifier with ISRUC-Sleep subgroup-III data for subject-1, 2, 5, 6,7,8,9 and 10 respectively. The highest precision was reported from subject-07 (97.26%). The specificity performance was achieved as the highest form subject-07 (92.71%). Furthermore, the F1-Score performance reports the highest value of 97.75% from subject-02. The kappa score for subject-02 and subject-07 is well suitable with SVM classification models. A comparative system performance with respect to other sleep studies has also been performed by the authors and is presented in Table 6.

Table 6. Result analysis in between the proposed method with previously contributed research works
Full size table

5 Conclusion and Future Directions

Maintaining proper sleep quality is crucial for both physical and mental health. The best way of measuring sleep quality is to classify the sleep stages using the single-modal signal. However manual inspection of sleep stages classification is quite difficult, consumes more time, and is highly subject-oriented. To overcome these difficulties, an automated sleep staging system was proposed which considers the basic steps such as pre-processing of signals, feature extraction, feature selection, and classification. Proper feature extraction and selecting the most suitable features play an important role during the automated sleep staging process and it ultimately affects the classification process. Besides a good amount of work has already been contributed in this field, but still, some gaps need to be dealt with. In this work, an attempt was made to address a few of these issues by presenting appropriate solutions for feature combinations, feature screening, and generalized classification models. Additionally, we also analyzed the effectiveness of the proposed methodology via various evaluation metrics. The following points are summarized from the proposed sleep staging classification results. The most significant features are chosen to improve the sleep staging classification performance compared to existing contributions and other methods considered in this research study.